Things are heating up: a real update in real-time AI video, some Seedance news, a look at Martini’s new camera control, and an experimental workflow worth stealing. The full rundown’s in the video.
Lucy 2.5: real-time restyling, actually affordable
Decart’s Lucy 2.5 works on live video at 1080p, 30 frames a second, with near-zero latency. One language gripe first: this gets called “editing video in real time,” and it isn’t editing in any non-linear sense — it’s doing the effects and compositing work on top of the stream. No less impressive, but a few of us bristle at the word.
What it does do: swap a character, restyle the scene, drop in VFX, or remove objects mid-stream, driven by text prompts and reference images. The trick is what Decart calls self-anchoring — the model adapts to its own generated output as the new reference point, which is how you avoid everything devolving into blobby face people or insane psychedelic nightmares. (Are there sane psychedelic nightmares? Let me know in the comments.)
In the playground I became an animated character, tried the hair-on-fire preset (I did need a haircut anyway), and then referenced in Renfield with “change the man to the attached reference character” — and there I was, Renfield. My note from the demo: I’ve got to work on how I act in AI restylizations.
There’s no app version yet — the phone-in-hand demo from their launch video doesn’t exist as a product — but the API is live, and at 720p it appears to run about 2 cents a second. With generation costs skyrocketing everywhere, that’s refreshing. It also means someone brave could build the app themselves. I’d give it a shot, but I’m super low on Fable 5 tokens and still don’t have Kimi K3 access.
The use case I keep coming back to is live previs: shoot on your phone, restyle in real time to see the shot, then run the final pass through a video-to-video model like Seedance for the render. That loop is now cheap enough to be practical.
Seedance 2.5 watch, Blomkamp, and a rumor
Still no release date, but samples keep trickling out. The latest, posted on Bilibili, is a car-crash shot with a great late-’70s, early-’80s Blues Brothers energy. The crash itself looks really good. Nitpicks: a phantom car appears at one point, some water sprays up from nowhere, and by real-world physics the trailing car would absolutely have piled in. Splitting hairs, maybe — but the nitpicks might be exactly why it keeps getting delayed.
Meanwhile, Neill Blomkamp — District 9, Elysium, Chappie, and to me the greatest Alien movie never made — released a short film created with Seedance 2.0. Link’s in the video description.
And one to file under nothingburger, but worth putting out there: I’m hearing more about a Seedance-level model waiting in the wings. I’ve spoken with a few people who say they’ve seen it; they can’t say much beyond that it’s very good and should be here soon. Given the growing concern about ByteDance dominating everything, some competition would help — with pricing, if nothing else.
Also from ByteDance: SeedAudio got an update — still 1.0, not a 1.1, but it now does timing control, including pinning dialogue to exact times. More control is the whole story there. Remember, kids: sound is half your picture.
Martini’s camera motion (sponsored)
This part of the video is sponsored by Martini — who are, as always, super chill, which is also the best way to drink one. No hard sell.
The big new feature is camera motion, building on their step-into-set feature: open an image on the canvas, step into it, and it becomes a Gaussian-splat-style 3D space. You set your lens length, fly the camera with WASD and the mouse, and drop keyframes on a timeline — multiple keyframes if you want to get wacky with it. Save the camera move and it lands back on the canvas as a motion input alongside your start frame, elements, audio (you can feed it a SeedAudio file), and prompt, generating in Seedance 2.0. My test was a camera move that makes no sense and a prompt that just says “she’s laughing,” which is exactly what I got.
The point isn’t my silly test — it’s the workflow it opens: plotting your camera move before your virtual actors ever step on the virtual set. A phone-as-motion-controller mode is coming, which should be great.
The quieter feature is Libraries: since Martini is canvas-based, you can now build a library — Flamethrower Girl, locations, whatever — and access it from any canvas in any project. And if you’re into MCP, you can pop into Claude and say “grab the Flamethrower Girl image from my library and place it on a new project’s canvas,” and the little minion does its job. Seems minor now; once the library gets substantial, it won’t be. Martini’s offering viewers 50% off the first month.
Depth maps as storyboards
An experimental technique from Josh Daws: instead of generated storyboard frames, use a nine-panel grid of depth maps as your storyboard layout, referencing your characters plus one image that defines the look. The result held a great late-’80s dark-fantasy vibe — Henson, Dark Crystal, by way of a sneaky dark-elf archer.
One important caveat: those panels aren’t actual depth maps — they’re GPT Image’s idea of depth maps, the same way Nano Banana will hand you something that’s kind of a Blender UI and kind of not. So I tested with a real one, generated from Theoretically Pose V4 — the free pose/depth tool I built a few videos back — using the same prompt and references, with the depth map as the video reference. The difference is subtle until you compare back to back: foreground characters separate from the background much more cleanly, and the palette shifts to the shot’s benefit.
I wanted to push further, but Seedance has been acting funky lately — most likely GPU allocation for 2.5 testing — so I’m benching this experiment until 2.5 lands. In the meantime: give the depth map thing a shot.
Final admin note: yes, I’m aware I need a haircut. Busy week ahead, so it’s not happening. Mostly what I’m saying is: see you again very soon.